I see a large IPM on the horizon. You're surrounded by journalists. Hold your position.
Let's just roll. We are surrounded by journalists. Hold your position. Get up. Trust the experts here. Five. We are experts.
Founder Vood. I see multiple journalists on the horizon. Stand by. UAV online.
Blaze. Double blaze. Triple blaze. Double kill. Fight. Go. Please. Team deathmatch. We are experts. Triple blades. That's just wrong. Right. Clearing order inbound.
We are surrounded by journalists. Hold your position. Strike two. Activate golden retriever mode. Marky, clearing order inbound. Five pudding. I see multiple journalists on the horizon. Stand by. Push that every time.
You're watching TBPN. Today is Wednesday, March 18th, 2026. We are live from the TBPN Ultradome, the temple of technology, the fortress of finance. Let me tell you about Ramp.com. Time is money. Save both. Ramp unites corporate cards, bill pay, accounting, and a whole lot more all in one place. Let's also pull up the Linear lineup. Of course, it's the system for modern software development. 70% of enterprise workspaces on Linear are using agents. We have a great show lined up. We have Matt Wang from Paradigm joining at the end of the show. Cameron from Maximus is coming in person. We got Andy Fang from DoorDash. Matt Multip. We got a great lightning round. Chris from Two Chrises. We're going back-to-back Chrises today. Anyway, thank you for tuning in. Thank you for watching TBPN as always.
The big news of the day, more news out of Nvidia GTC. Lots of Nvidia announcements. The stock is up. It's a $4.44 trillion company last time I checked. That is big. Nvidia has been on an absolute tear. And some really promising things with the Groq acquisition already, or like pseudo acquisition, the deal, the Groq partnership. They're already starting to explain a little bit more about how those two technologies fit together. And there's some good coverage in Stratechery about that. But the big news out of Nvidia yesterday was that Nvidia says it's restarting production of AI chips for sale in China specifically. So, Jensen Huang says the company's supply chain is fired up after months of mixed signals from the Chinese market. We've been tracking this for a long time. Of course, chips were banned from sale to China in 2022 by Joe Biden under the CHIPS Act. That also unlocked billions of dollars in incentives for American chip manufacturing. We began following the Intel story. We began following the Huawei story. There were a number of different initiatives. And then the narrative flipped back and forth, back and forth on what are the risks and what are the costs and benefits of actually selling chips to China. Back in 2022 after the CHIPS Act, which if you want to read up more on it, we've interviewed Chris Miller, the author of Chip War. It's a great book. Highly recommend it.
We should have him back on.
Yeah. Ben Thompson's also covered the story throughout. There's a whole bunch of good stuff in Stratechery on chips in China. And the calculus has always been pretty clear, like on the first pass, the first-order effects. AI is an important technology. America wants an advantage in the AI race, the AI buildout. So less chips for China means more chips for America, stronger US economic engine.
Massive chip shortage right now. You see old chips being valued basically more than they were when they launched, which implies there's plenty of demand in America, so you'd want to keep those here, but...
And plenty of demand at TSMC from American chip makers specifically. I mean, even Apple is sort of getting crowded down there on the 2nm, which I think is better for phones than for GPUs, but they're grappling with what the compute boom, what the AI buildout boom will be for their business. But the second-order effects are what are more complicated. So, in November of 2022, I was writing about this and I said on a YouTube video, in the past, China's dependence on foreign technology companies has been seen as a key bargaining chip. Why would China invade Taiwan when they need to keep TSMC's manufacturing facilities online? Chip manufacturing is extremely precise. One of these factories isn't going to withstand a rocket hitting it. So if there is any sort of broad military action in Taiwan, TSMC probably gets a little bit damaged. I mean, even the tiniest earthquake...
They have to track the weather outside of the facility. If the weather outside of the facilities is fluctuating at all inside, earthquakes, Taiwan is famous, or TSMC is famous, where they don't even need apparently a detection system or a push notification to their employees if there's an earthquake. If the employees sense that there's an earthquake, they just get up and go to the factory and start working on things. It's not like they need to like, oh, we need to email all the employees. The employees just know because that's the level of sensitivity over the TSMC fab.
So if you want to keep TSMC producing chips, you can't invade Taiwan. And so by banning the export of chips to China, the cost to China of a Taiwan invasion decreases. And so if China can't access TSMC chips anyway, it's a lot less risky to go to war. And that was always the risk with hardcore chip bans, that you'd create an increased risk of geopolitical conflict. And in 2022, the Russia-Ukraine war was about six months old. Global conflicts have grown significantly since then. Obviously, we've been tracking the Iran war. And America's military is potentially stretched thin. So, the risks of a Taiwan conflict are higher than ever. And so you add to the fact that everyone agrees that we will be in chip constraint, the chip shortage through at least 2030. And the need to keep TSMC supplying chips to American companies is extremely important. It's always been difficult to parse the various arguments around selling chips to China because there's an insane amount of money at stake and many, many people whose basically their full-time job is to advocate for a particular position. And I was talking to a friend who has interfaced with a lot of the different groups in Washington, and there are like huge think tanks, huge lobbying organizations on both sides, because there's so many different interests going on.
Yeah. Not to mention how much of everyone's retirement accounts Nvidia actually makes up.
Holding up the world economy, right? That's the meme. And that's certainly true. So there are good arguments on both sides. One that keeps getting trotted out is, you know, it's important to keep China dependent on the American AI stack, and it's reasonable. The better argument might just be dependent on a functional TSMC fab, but there are benefits to the CUDA ecosystem and to the idea that whatever models get built there will be applicable here. We'll be able to transfer that research and development that happens over there very quickly. But the much more important thing is just that the more the economies are interlinked, the less likely there is a conflict. So all of this underscores the importance of TSMC Arizona, Samsung, Intel broadly, as well as startup fab projects like the TerraFab, and there's a whole bunch of other companies that are at least talking about tool making, talking about getting a fab up and running. Nothing that's really taken off. I mean, even Samsung is not fabbing any of the leading-edge Nvidia chips at this point. But, you know, there's incredible economic incentive, and people have known that this is an important sort of cash cow business for a few years now at the very least. And so there's certainly been a lot of incentive to put plans in motion that might play out before 2030 because there's such a shortage. And so in general, we're traveling this long and narrow road. But I'm coming around to the idea that selling some chips to China is the best possible move at this particular moment in time. It was easy to stick with the first-order logic of just we want the chips because we can use them to power our economy, so we should have them. But going forward, it'll be really interesting to track how China's indigenization project, their domestic supply chain, evolves. They, of course, have...
Yeah. My take has always been like, even if they are getting chips, it's not like the party is going to say, actually, our domestic supply chain is no longer important because we're getting a drip of H200s. They're still going to keep the momentum that they have. It just wouldn't be like them.
There's a lot of debate over that because momentum comes from...
You can take some of the wind out of the sails. Yes.
But they can just spend more...
Slow the momentum potentially.
That's the argument, is that by limiting demand, you like, there's a local fab in China, a chipmaker, that says, okay, well, you know, our demand is half as much, so we can't afford to scale. Sure, we have the money, but we don't really need to deliver this because there's no buyer. So you don't get the process level of execution, you don't get the excellence that comes from actually needing to run the real business. It becomes more of like NASA than SpaceX. That's always the risk with like, you know, you're just throwing government money after it. Before we move back to the Wall Street Journal, let me tell you about Figma. No matter where your idea starts, whether it's Claude Code or Codex or Sketch, the Figma canvas is where ideas take shape and products, ideas connect and products take shape. Build in the right direction with Figma. And let me also tell you about Public Investing for those who take it seriously. Stocks, options, bonds, crypto, treasuries, and more with great customer service.
So, Nvidia and the Trump administration have been involved in a complicated tango over the sales of its advanced artificial intelligence chips in China. Last April, the Commerce Department halted exports of the H20, a processor Nvidia designed explicitly for the China market. That was the nerfed H200 that was not supposed to be able to train as advanced of models, but it was reported that the high-flying team behind DeepSeek sort of figured out how to use those chips effectively. So there was a debate over is the H20 actually just as useful as the H200 or close to it or closer. Well, it's sort of a moot question now because H200 is coming to China, which is the more advanced version, the not nerfed version. So the company's fortunes turned again in December when the US said it would allow Nvidia to sell its H200 processor, a chip that is a generation behind its most powerful S-series of GPUs. So this is not the B200, not the Blackwell, this is the Hopper, but it's still very advanced. In China, as long as the company shared 25% of its sales with the US government, so there's basically an export tariff. GPUs or graphics processing units are powerful chips used in AI training AI models. I think everyone knows this. In late January, after Huang visited China, officials there signaled they would approve H200 sales as well. So there was a big debate over whether or not Beijing would block the imports of H200s in order to stimulate the indigenization of the domestic semiconductor supply chain. It seems like these will be going through and they will be purchased.
Inspire local industry to grind harder.
Basically. But you have a whole bunch of AI labs that are not in the fab industry. They're not involved in the verticalization process, and they just want to train great models because they have real businesses. Like they're running Alibaba or Baidu or any number of companies, and ByteDance wants good AI and they want the best chips, and they're not really interested in getting into the supply chain at a deep level, at least in the short term. So, but until Tuesday, the status of the H200 in China was unclear. In Nvidia's most recent earnings report, the company said that although it had received approval to ship small amounts of H200 products to China to date, we have not generated any revenue from those sales. Speaking at the company's GTC event Tuesday, Huang said that in recent weeks, demand signals out of China have strengthened. We have been licensed for many customers in China. We've received purchase orders from many customers, and we're in the process of restarting our manufacturing. Our supply chain is getting fired up. Nvidia didn't comment on how much it expected to earn from H200 sales in China, but in the past, the company has said that the Chinese market for its AI processors could be worth tens of billions of dollars a year. That makes perfect sense.
And Christina over at CNBC got the original scoop.
Near the bathroom line at GTC, Huang happened to walk by security and asked if he'd answered her earlier question. She said she had one more.
You said you got China purchase orders. That means you got the green light from both sides. He said yes.
As a scoop athlete, you never know where your next scoop is going to come from. You got to be ready.
The sticking a microphone in someone's face as they're walking by really does make sense.
Yes. Yes. Jensen loves answering some questions on a microphone line. Chinese authorities approve Nvidia's H200 chip sales, source says. And TOR taxes says Xi Jinping has given up. The situation with the H200 sale in China summarized, and it's everyone jumping off and Xi Jinping saying, stop, domestic chips can work maybe, and everyone at the lab saying, must get CUDA, dude, Nvidia is the future, H200s, and they're falling into the CUDA dependency mode. It will be interesting to see the evolution of the CUDA mode as other chips come online. I mean, we've seen TPU become functional. So there are questions about the CUDA long-term, but it certainly seems like not a stopper to demand right now as labs place orders from Nvidia.
In other news, J.R.R. Tolkien used Gen Z brain rot slang over 70 years ago. That's how ahead.
He was in the quote, 'Maggots, ye're cooked.'
White skins will catch you and eat you. They're coming.
Oh, he's like using it literally, like you will be cooked by some, I don't know, villain, I suppose.
A cry from the Gishnok showed that this was not mere jest. Horsemen riding very slowly had indeed been sighted, still far behind but gaining on the orcs. So the orcs will cook you if you fall behind, I suppose. Let me tell you about Okta. Okta helps you assign every AI agent a trusted identity so you get the power of AI without the risk. Secure every agent. Secure any agent with Okta. And let me also tell you about Lambda. Lambda is the superintelligence cloud. Building AI supercomputers for training and inference that scale from one GPU to hundreds of thousands. Nvidia's biggest GTC announcement was a $20 billion bet on the same problem that Cerebras solved six years ago, says Andrew Feldman, the CEO and founder of Cerebras.
Shots fired indeed. He says, 'Their next-gen inference chip, not available yet, has 140 times less memory and less memory bandwidth than Cerebras. To run a single two-trillion-parameter model, you need 2,000 Groq chips. On Cerebras, that's just over 20 wafers. Even paired with GPUs, Groq maxes out at a thousand tokens per second. We run at thousands of tokens per second today and every day in production now. Why? When you connect 2,000 chips together, every interconnect has latency. Every cable has overhead. It doesn't matter what your memory bandwidth is on paper if you're bottlenecked by the wiring between the thousands of tiny chips. We solved this with wafer-scale, one integrated circuit, no interconnect tax. Jensen told the world that fast inference is where the value is. He's right. It's why the world's leading AI companies and hyperscalers are choosing Cerebras.' And so he puts up a little graphic of Cerebras versus Rubin plus Groq together on one system, and he is touting 90 times the amount of memory, 90 times the number of chips needed to run a two-trillion-parameter model. I don't know how relevant. I wonder how this scales relative to, well, what if you're not running a two-trillion-parameter model? If you're, you know, five times lower, does this number get much more manageable? It does feel like he, you know, I mean, he's talking his book here. But I don't know, there probably...
It is a lot to talk about. It's a big chip.
I don't know. I mean, have you, Tyler, have you ever run into any Groq chips in the wild? I remember the original demo was inferencing Llama, that was pretty fast, and everyone was like, this is amazing.
I've hit the API a few times. Okay. Because yeah, you can like, there's just like consumer API where you can run like local smaller models.
I had heard you need a lot of chips, but that's not on its face like a huge problem necessarily. But it did seem like they had made some design decisions that maybe painted them into some particular corners, but those corners might be still economically advantageous in certain cases. But...
Yeah, I mean, it was definitely still slower than there was like the ChatGPT thing, right? Because that's very much like, it's literally the model being served.
Well, I mean, do we know that that's what's happening with the ChatGPT demo? Because like the Groq demo was very much like, we have a chip, here's a demo, and it was actually 2,000 chips behind the scenes or a bunch of chips together behind the scenes. And so it's possible that the ChatGPT is doing something similar where in order to speed things up...
We don't know. I just asked ChatGPT, 'What's your favorite thing about Tyler Cowen from TBPN?' Couldn't find any information about a person named Tyler, especially in the context of TBN, which might refer to The Big Bang Theory, a popular TV show, or maybe The Bachelor...
The Bachelor Paradise Network. TBP. Okay. Yeah, that works.
It's fast, but it doesn't... I don't think ChatGPT has search. Yeah. Like I think it's all just trying to memorize things, which would slow things down.
Let's play this video of... I want to play this video of Jensen back in September 2009. Bubble Boy is having some criticism of GTC because apparently people are getting up and asking questions that are intended to pump bags. He says, 'GTC has turned into a conference less about tech and innovation and more about pumping your bags by getting a Jensen soundbite on some niche supply chain player.' And so someone stood up and asked, 'Hey, why'd you invest four billion in these companies?' I mean, that's a reason.
Well, let's take a walk through history. I make content for... so they can have the average American and people around the world can have a...
Let's play this Jensen clip from 2009, September 2009. Jensen walks onto a small stage at the Fairmont Hotel in San Jose.
About 1,500 people in computer science. You know, if Dave Patterson, John Hennessy, and I know that Pat Harrington is sitting here watching me, you guys are going to go nuts.
You need to put a mic on him.
This is called CEO. This is not computer architect now. This is called CEO, but it illustrates the point.
Well, it makes a lot of sense. This is an illustration of co-processing.
Yeah, Carlos says no jacket.
I realize there's a lot of words.
No jacket. He is looking pretty simple.
We're going to do a test at the end of this presentation. So this is a simple map. The first section on top is an example of a program that is parallel-intensive, and on the bottom is an example of a program doing camera...
And the way that I simply made it is the serial code part of the parallel-intensive program takes about one second to run, whereas the parallelizable part of the code takes 199 seconds to run, which makes sense. It's an enormous amount of data.
Nvidia was a $7 billion company when this happened.
Jensen kept saying the GPU wasn't just a gaming chip. It was a computing platform. Kept saying parallel processing would reshape every industry from medicine to finance.
Yeah, yeah, yeah. Just put the video games in the bag.
Seriously? Yeah. This was, yeah, before training neural networks. Wow. Pretty remarkable. 17,000 people are at GTC 2026 now, packing a hockey stadium, watching the same guy explain what comes next. Very, very interesting. Anyway, let me tell you about MongoDB. What's the only thing faster than the AI market? Your business on MongoDB. Don't just build AI, own the data platform that powers it. LinkedIn...
SemiAnalysis is over there slopping it up. They say LinkedIn is the only social platform where you can post a sloppy reaction poll and it will get tons of engagement.
Wow. They're just really throwing shots.
Give LinkedIn a chance on a long enough timeline. They will all come to X.
It might take 30, 40 years. They'll make it over to the dive bar eventually.
I like that SemiAnalysis leans into the particular rough edges of every platform. Like I follow them on Instagram and they actually post just like hilarious vibe reels and brain rot.
It's just like brain rot.
Yeah. But often it's completely... I'm the only one that's liking them. I comment on a lot of the videos.
Yeah, I comment too. It's amazing. Like they are not having broad success. But as far as like what Instagram...
Well, it's like, you know, the fine details of, you know, inference max or something, and then it's like Minecraft parkour over it.
Yeah. I was in a few of them because they clipped our interview with Dylan Patel from Cisco AI Summit and put some videos on there. But there's a hilarious one where it leads in with some text about, like, you know, SemiAnalysis, semiconductor analyst, like explains this thing, and then there's just no audio while he's explaining it, and then it just cuts to me, and I'm like, and then it cuts back because we were editing the footage. And so I don't even know if it was like a mistake upload or it's like intentional, but it's all very, very funny.
Apple cracks down on vibe coding apps. It's over for you, Tyler.
Apple's moves come as vibe coding apps help people create apps.
Vibe coding apps, not vibe-coded apps.
I know it's over for him times two.
No, because he doesn't vibe code on a phone.
I'm kidding. I'm kidding. I'm kidding.
He doesn't vibe code on a phone.
I'm kidding. But Apple's move comes as vibe coding apps help people create apps for Apple devices as well as web apps that aren't listed in the App Store. Let's get into Stephanie and Aaron's reporting from The Information. Apple has quietly prevented AI vibe coding apps such as Replit and Vibe Code, which help people create games and other applications, from releasing updates to their mobile apps on the App Store unless they make modifications. The company confirmed it has told some app developers that the vibe coding capabilities violate long-standing App Store rules that say an app can't run code that changes the way it or other apps function. Apple's crackdown is happening at a time when vibe coding apps are emerging as a potential threat to the company by helping developers create web apps that aren't listed on its App Store, a key source of revenue and profits for Apple. Some of these vibe coding apps also help developers create apps for Apple devices. That ability has likely contributed to the explosion of new apps launching on the App Store in recent months, leading to a slowdown in the approval process in some cases, developers say. An Apple spokesperson said the policy isn't specific to vibe coding apps. Following The Information's questions about the standoffs, two of the people with knowledge of the situation said they believed Apple was on the verge of approving updates to Replit and Vibe Code.
Those app makers had agreed to either tweak their apps the way their apps showed customer previews of vibe-coded apps or get rid of certain vibe coding capabilities entirely, like making apps for Apple devices. So it sounds like you're able to basically generate an app with Replit and then use a preview of it that maybe is functioning a little too much and effectively allowing Replit, the app, to do things that Apple didn't approve of. So...
Yeah, the App Store. That's what I was looking up. The App Store is always so fascinating. So if you search Vibe Code, you get an ad for Replit as the first response. Replit is number three in DevTools, has 14,000 five-star reviews. Then Vibe Code is listed as Vibe Code Website Builder, has 3.3K. Pretty solid. It doesn't look like it's charting, but it says, 'Learn how to vibe code. No experience needed. Build websites with professional designs.' Much more focused, I think, on static content. But there's been a number of these like website builders in the App Store for a very long time. Then Replit ranks number two when you search for Vibe Code because it says Replit Vibe Code Apps. Then you get Sticky, an AI game maker, Vibe Coder, AI app creator, Anything, the AI app builder, Vibe Code, Claude and Codex AI with 21 stars. There's a lot of people that are just jumping in. There's one that's gone.
Replit's argument is that they're helping the user generate an app that's just opening it up in a web view. So it's effectively just a mobile...
An app built for mobile web. And Apple's arguing that no, that's effectively in the Replit app.
Would you download this app, Jordi?
I don't know if you can see, but it's definitely vibe-coded the App Store preview.
It's like a graphic, like an AI image of the GigaChad using the computer. It's very funny. But man, there is so much IP infringement. Love a code. Not from Lovable. Love a code, build with Vibe Code. They're like, I wonder who they're trying to SEO against. Clearly there's a whole cohort that are just trying to steal market share from incumbents that have name recognition.
Do you remember that company that was doing vibe coding on the iPhone and they would tap your phone and basically AirDrop you the app? We talked to them at YC Demo Day last year. Yeah. And there's a number of these companies that are, you know, trying to be like the AI game store, sort of like the Meta Simulator, like build a simulator and create a harness that's really good at vibe coding a game. It feels like a really valuable category if you can crack it, but you are going to be bumping up against the App Store all...
You just have to compete with Sam Altman, Dario Amodei. Yeah, maybe I...
I do wonder if there will be some sort of... I mean, Roblox would be like the bigger one maybe, or does it come out of Codex and Claude Code instead? Like, how do you... I wonder what the barrier and how much distribution matters there for that audience of people that want to vibe code a game.
Yeah, for gaming, I just think Roblox is just going to continue to be like Roblox is the Roblox of vibe coding.
And yet, we did not build our simulators in Roblox.
Yeah, but they're not like massive multiplayer games.
And we want people to be able to click a link and use it on their phone immediately. Like we actually...
We have a new simulator coming, by the way.
We're addicted to simulation.
And this one we're putting a little bit more effort into. And John's already addicted.
I would say this one is like just actually fun. Jeremy Gohn Simulator was more just like a different way to experience.
It was for people that didn't just listen to his episode on investing, they simulated it. Yeah. But the actual interaction pattern of like guessing and whatnot, like it was like, oh, this is novel, and then it was like, okay, this is a chore, and then it was like, okay, how many questions are there? 200.
There were what, like 60 questions or 48.
This new simulation, you're not going to be able to put it down.
You're not going to be able to put it down.
And it can effectively solve some industry-wide issues that are happening right now.
But we're going for impact. We're going for fun.
But, you know, Apple's had this long-standing policy around do not, like, they want to review the software. And so you can't create an app that rewrites its software. But yeah, I wonder if Apple can do anything to create more like a mini peer-to-peer experience because I remember I was learning how to build iOS apps when I was an early teenager and I was so frustrated that I had built Pong, but I couldn't just share it with my dad and say, 'Hey, you can play this.' Like it just wasn't... Yeah, we can do TestFlight, right? Yeah, TestFlight, but TestFlight is still like, it's certainly not designed for... it's not like a social peer-to-peer experience. Like you still have to opt into the TestFlight network or do all these jumps. And for some reason it is weird that it kicks me out of the Apple ecosystem when I get a TestFlight...
Says high-key Tyler could make better Siri and Replit with a thousand-dollar budget. Better Siri and Replit. Two massive things.
No, better Siri in Replit.
Oh, better. Oh, okay. Oh, yes.
For his last thousand bucks. He's down to his last thousand bucks.
Yeah, maybe that should be the new constraint on hackathons. Instead of it being like two days, it's like twenty grand is the max budget. And what can you do with that budget for the token allocation?
His big budget for hackathon.
Yeah. I mean, maybe the $200 plan or something, like exhaust it. You can't use multiple or something like that.
The question is how much is Apple itself vibe coding because the software quality in the apps that I use...
Mark Gurman said they're using Claude all the time, right?
Yeah. No, but to me, I'm saying so far my experience recently... I've had an issue with the most important application on my phone, which is...
You were complaining about the Photos app.
That was just poor design.
The Phone app has gotten a lot... Yeah. John finally came around because the Phone app is like, you're like, okay, I'm going to hit this button. I might be calling this person out of the blue even though I just want to...
And I'm not sure what phone line I'm calling them on. I guess they're designing for a world where people only have one phone number. But I still have a lot of people. This is blowing your mind, Tyler. But back in the day...
Back in the day, people used to have multiple phone lines, multiple phone numbers. It's true. So, but yes...
Like a work phone. I mean, when I was at Facebook, I had two phones, but...
Two phones. That's right.
But a lot of people will have a home phone and a mobile phone. And so that was the thing that you saved in your contact book a long time. And the problem with the new iOS Phone app is that I've been calling randomly people on their home phone if I have it saved. So I need to maybe go delete those numbers or put them in a comment field so that it always calls their iPhone because I have moved to just calling people on their mobile phones. But anyway...
Well, the good news, we have our first ever Apple employee coming on the show.
Really? Oh yes, we do. We do. I'm very...
That's going to be exciting.
Apple employee on the Technology Business Programming Network.
That's not... I mean, a current Apple employee. We've had ex-Apple employees, of course. But quickly before we move on, let me tell you about Fin.ai, the number one AI agent for customer service. If you want AI to handle your customer support, go to fin.ai. And let me also tell you about Console. Console builds AI agents that automates 70% of IT, HR, and finance support, giving employees instant resolution for access requests, password resets. So, Martin...
We got to go over to SF with Martin.
Yeah. Wait. Oh, okay. This is part two. We're just jumping straight into this.
We're going straight into part two.
Okay, we're going straight into part two. Take...
He says, 'Have you ever had the thing that you know a lot about become the current thing?' That's happening now with peptides.
Holy [__]. I don't know where to start. Pharma basics. Most people obsessed with peptides don't know a few things. Peptides as pharmaceuticals have been around since the 1950s.
A peptide is just a small protein. Peptides have...
have extremely short half-lives.
1950. That's like 10 minutes before 8:00 p.m. yesterday.
Yes. Correct. Peptides have extremely short half-lives, often on the order of seconds or minutes. So if you're saying you're interested in peptides, you're saying I'm interested in biopharmaceuticals, but only drugs with very weak pharma...
Pharmacokinetics. That's a new word for me. Drugs, of which peptides are a subgroup, usually have a specified target. This is an electrostatic interaction, usually hydrogen bonding between the atoms of the drug and the atoms of the target, typically, but far from always, a receptor. If you can't tell me what the target is and how the drug is binding to it, you do not have a drug. You have delusion. Next, drugs are rigorously tested, not only for safety reasons, just identifying the pharmacokinetics of a substance, how it travels in the body.
Pharmacokinetics, a collab...
They're linking and building.
is arguably the most important starting point for any medicine. How is it metabolized? What is its half-life? Without this basic information, you can't even begin to have a medicine.
You can start pharmacokinetics in animals and scale to humans, but you also need a therapeutic hypothesis. This is a thoroughly vetted biological idea considered a priority as to why this medicine just might work. You very rarely discover these after the fact. Determining target engagement requires assays.
This is going to be a rough one.
Brutal. A priori, exposed.
What assay was your drug tested in? What did it show? Direct target engagement is very important to falsify your biological hypothesis. And you can continue, John.
Preclinical studies are so manufactured and fraudulent in today's day that I wouldn't rely on them for biological hypotheses unless they are from an incredible lab, were done a priori, etc.
There we go. Clinical reality is far harsher. Without a double-blind placebo-controlled study, there is often nothing to talk about. If I hear 'but I know dozens of people' one more time. Exclamation point. Exclamation point. Exclamation point. Screw the FDA and pharma. Really? Really? Coming from Martin Shkreli. He is saying maybe don't screw the FDA and pharma. When most of the SF and elsewhere crowd talks about peptides, they're not thinking, and this is going to be a hard one for me, octreotide. They're thinking some random stuff that's been thrown in animal models and is not FDA approved. Look, I'm not a softy. If there was a drug that could help me or my family, I'd find a way to get it. But I'm also not stupid and spent 20 years looking at pharmaceuticals. Drug companies like to make money. Drug companies love looking at random molecules and putting them in clinical trials. There are thousands of biopharmaceutical companies that are publicly traded. It is not hard to do a clinical trial from a university. If your drug has never been tested, there is a reason. The reason is not that you are a biopharmaceutical genius who has found something cool that everyone else missed. The FDA plays an important role. They make sure that whatever is on the label is actually in the drug. That's why prescriptions are important. If I operated one of these research chemical shops, wildly illegal, I might add, I would just ship people alanine or something. No one would have any idea that it wasn't BPC-157 or whatever is popular right now. The other side of the argument. But Martin, there has to be some unapproved drug out there that's useful to take. Yes, there are plenty. That is how I made a living, says Martin Shkreli. But it is not for you, world traveler, to think about this. The things you know do not...
apply to pharmaceuticals. It's not that you're not smart. I'm sure you're smarter than I. It just takes practice and time to understand medicine. I believe some places will even require you to go to school before you can decide who takes what drug. Just ask your doctor for medical advice. There's a reason you don't do surgery on yourself, fly a plane by yourself, etc. But Martin, I want to optimize my health.
You could fly a 747 then.
I could. I agree. We don't need studies to know that you could land a 747 if I needed to.
Yeah. Oh, 100%. But Martin, I want to optimize my health. No, stop it. You're not sick. It's all nonsense. Leave medicine to physicians. You do not know what you are doing. Become a physician if you are that interested, or spend a lot of time and money on biopharma. I have zero doubt you'll change your mind. There are no health care professionals that I know of who give an [__] about these unapproved research chemicals. But there are actual dying people in the world. Duchenne muscular dystrophy, pick a disease, go fix those diseases. You'll make someone and their family a lot happier than blaring that you know about medicine. This has to end.
So we are going to have a debate. Max Marion from Superpower will be coming on Monday at 12 to debate Martin Shkreli, the professor himself.
So, we're going to have a little debate. Superpower, I believe, sells these small proteins.
And so, it'll be an interesting conversation. So, Monday at 12 Pacific, we can look forward to the great debate. So there's some conversation in the comments here. Michael Dugan says when people talk about peptides they mostly mean things like retatrutide, which is in stage three clinical trials and looking extremely good, or BPC-157, which has tons of clinical and anecdotal evidence. Your critique is just self-aggrandizing fluff that falls apart when you apply it to the actual examples most people are using. So he's saying, 'Look, most people aren't using this stuff that's like crazy far out there. They're just pulling forward things that are actively being worked on by the pharmacy.'
Just another pod guy says, 'Great post. Would be even better if you were in a flow state with a low dose of reta.'
Martin does not think he doesn't like BPC-157. He says BPC-157 has no evidence. LMAO, reta is literally a biopharmaceutical from Eli Lilly. The concern with BPC-157 has always been that it could accelerate cancer growth.
And so because it hasn't been studied well enough in humans, that is a risk that people I think should be aware of.
How much of this is actually because of people's AGI timelines? Is there a real overlap in San Francisco between like, yes, it might give me cancer in 20 years, but I think we will cure cancer in 10. So if it makes me look good in the next five...
But I think it's all about people just want some type of edge. They want to alter their state. It's somewhat basically human nature.
You can make the same argument for why you should wait though, because then in 10 years...
AGI will create like a super drug that would just instantly make me jacked.
Oh, right. It's like you can do it either way. No one has a decade to wait to be...
Yeah. You want to be jacked now.
Because all things equal, if the cancer risk of both scenarios is zero, you'd rather be jacked for 45 years as opposed to 40.
How much would we have to pay you a day to not lift anything heavier than a single piece of paper?
There's no amount of money.
Ridiculous. Let me tell you about Turbo Puffer. Serverless vector and full-text search. Built from first principles in object storage. Fast, 10x cheaper, and extremely scalable. And let me also tell you about Cisco. Unlock critical infrastructure for the AI era. Seamless real-time experiences. New value are unlocked every day with Cisco. Pretty iconic. Someone hire this kid. Robin Smith is sharing a poster.
Warning. Warning. This man has been known to increase shareholder value with a QR code out to his LinkedIn.
And his name is Lachlan von Eggman. That's a great name. And he's looking for a job as a summer 2026 electrical engineering intern. And I cannot recommend this gentleman enough. He has the correct mindset to go into a summer internship optimizing.
Lachlan von Eggman, you should get a job as an electrical engineering intern. But if you don't, hit us up.
Yes, this is good marketing. Yeah, you can.
Maybe we could hire Ben. Maybe we could hire this guy to work on our little electrical engineering problem back here. We got a lot of cords.
Doing cable management. I think he has bigger fish. But whatever increases shareholder value, he might go for it.
If you can handle cable management in the Ultra Dome, yeah, you can go far. Dependable thermonuclear work ethic. That's a good phrase, thermonuclear work ethic. He says he's smart with a capital S. And he shares his email. So if you're looking for an electrical engineering intern this summer, give Lachlan a call or shoot him a...
head over to talking about Eric Seufert, okay, at Mobile Dev Memo. He writes, 'Private credit and the AI value reallocation' starts with a quote: 'The essence of technology is in a lofty sense ambiguous. Such ambiguity points to the mystery of all revealing, i.e., of truth.'
Martin Heidegger, I like it.
says a familiar intuition about the emergence of any new technological paradigm is that new methods of engaging with the world create uncertainty which is most easily interpreted through the lens of perceived negative outcomes. As private credit markets deteriorate, it's tempting not only to blame AI for that decline, but also to extrapolate any dislocation to its logical extreme. That is where rising default expectations among software companies are increasingly framed as early signs of a global systemic crisis. The most convenient analogy is the GFC.
And yesterday, Carried No Interest was giving a little bit of a doomer take around some of these software private equity deals, but he was not ringing alarm bells to the tune of the global financial crisis. Correct. He was just saying that some of these deals are underwater. Some of these investment professionals might be needing to join different firms to find different opportunities. Sort of the bull case for special situations, right?
Yeah. Leaving the firm and be like, 'I didn't really work on it. I was an investor, but I didn't do much investing during 2018 to 2022. I was mostly just sitting there saying, Guys, I don't think we should do this deal.' The beauty of private credit is that you have all these different funds that are being deployed, that have been deployed on different time horizons, that have longer time horizons in general, right?
And so you can have basically like a rolling collapse versus like a run on the bank where you have like one day where everyone realizes...
It's kind of like the worm. Like you were doing the worm yesterday.
Yeah. That does not seem great. It's like a wave starts small and it gets bigger and bigger and bigger, right?
No. Yeah. No. So, you're 68. You're starting the worm and you're really tall and you're kind of coming down a tree falling, but then you bounce back up like this and you're steadily kind of like losing. I hope that's not where this goes.
Let's continue. This framing is incomplete, says Eric Seufert in Mobile Dev Memo. He says it isolates the destructive effects of AI while ignoring the mechanisms through which those effects propagate and where value ultimately occurs. In this piece, he makes the case that contemporary economic conditions bear no resemblance to those leading up to the global financial crisis of 2008. I also argue, he says, that any weakening in various categories of the software landscape as a result of AI will not only mostly remain contained there but will likely lead to economically expansionary productivity gains and efficiencies that offset potentially disproportionate losses in private credit. That's very exciting. The private credit market, he gives a little history on the private credit market. The private credit market has grown precipitously since the 2020 era co-pandemic. In a note from the Federal Reserve, you're just like, there will be a rolling collapse, but also I'm excited to grow.
You got to hit the air horn for growth.
In a note from the Federal Reserve, the author's remark that...
It is funny that if you go back to our early bits, yes, we were joking about praying for bubbles, encouraging leverage.
People were listening. I'm not saying encouraging ads.
Yeah, a lot of things came true.
Obviously, we were joking. They were bits. I guess some people they've taken it a little too seriously.
Isn't there some humor in every joke? Isn't that the point of humor? You mean some truth?
Some truth. Some truth in every... What do I say? Some joke.
There is some humor in every joke.
Yeah, some truth in every...
John Coogan once said, 'Wow, there's humor in every joke.'
I'm only on my second diet joke. Okay. Don't talk to me. Don't talk to me until I've had my 10th energy drink.
Anyway, this is what the Fed had to say. Private credit has emerged as one of the fastest growing segments of non-bank financial intermediaries, NBFIs, over the past 15 years or so, reaching a total asset class size of 1.34 trillion in the US alone by the middle of 2024. A report from Morgan Stanley published in October of 2025 estimated the size of the private credit market at the start of 2025 at three trillion. A report from Vanguard said it was 1.8. So people are sort of all over the place. Private credit is an asset class that functions as a parallel banking system. Don't call it a shadow banking system. It's just merely parallel in the darkness.
Look, I'm gonna be concerned. Here's what I'm gonna be concerned. And you have leaders at these private credit firms that say, you know, look, private credit's amazing. Yeah. But it's unfair that we're keeping all these gains private and we need to make them public.
Public. Yeah. Yeah. Yeah. So some type of federal kind of involvement in the private credit sector could make sense.
It would also be really, really bad if like one of the most respected leaders of one of the biggest banks in the world was to compare the industry to like a cockroach. Yeah. Like some really like, you know, some bug that would be...
You don't want to be compared to a scuttling creature.
No, no, no. Maybe a soaring eagle instead.
There are some soaring eagles in this portfolio of private credit assets. So while it predates the global financial crisis of 2008, the scale of assets under management for private credit began growing precipitously following the global financial crisis to fill a funding gap. As the regulatory standards for lending and capital requirements applied to banks, such as those introduced by Basel 3, intensified in the wake of the GFC, non-banks such as private credit lenders became a key source of funding for companies, especially those operating unprofitably, as many software companies might be earlier in their life cycles. Chernenko specifically attributes an increase in non-bank lending to regulatory rigor. So the banks become more regulated after the financial crisis and so non-bank lending that is not subject to the same regulations increases in many, many such cases. Another note from the Federal Reserve identifies private equity as a significant vehicle through which private credit is deployed because private equity acquisitions are often funded with debt through leveraged buyouts. They used to go to banks for those lines of debt. Now they're going to private credit firms.
Okay. And I think we should fast forward to down further private credit and the AI bull case because this article is very long and you should go subscribe to Mobile Dev Memo and read it entirely. But let's skip to the AI bull case.
Okay. So last month, Blue Owl sold $1.4 billion in private loans across three of its funds and announced that it would end regular quarterly liquidity payments in one of its funds, OBDC2, opting instead for payouts tied to asset sales and other liquidity events. This month, Blackstone experienced a record number of redemption requests from... they wanted to inspire their LPs to grind harder.
The company indicated that it would honor all requests.
If we give you too much mailbox money...
Mailbox money? I've never heard that term before.
Mailbox money is just money that you know like shows up. You have some apartment somewhere and they just send you the check.
Okay. No, I've never heard that before. Huh. Wait, do you know that, Tyler? Mailbox money?
Am I crazy? I don't know.
Mailbox money refers to passive income that arrives regularly with minimum active effort, commonly in the form of real estate, oil and gas royalties, or dividends.
That's why you listen to TBPN.
It also is a mixtape by Nipsey Hussle.
Nice. That's why you know it.
Also this month, JP Morgan marked down the value of loans the bank holds as collateral from private credit funds related to software businesses, reducing those funds' ability to borrow. Some of this negative sentiment is attributed to recent bankruptcies from firms to which private credit facilities were exposed. First Brands and Tricolor are among the headline cases. But the substance of the current negative outlook might be traced to fears that AI will render the vertical SaaS products into which PE funds invested in the 2020 to 2022 vintage irrelevant. So this is the SaaS apocalypse narrative but playing out in the private markets. In a recent note, JP Morgan states, 'Private credit has 21% exposure to software, driven by the attractiveness of SaaS recurring revenue models and the fact that 96% of software companies are privately held. Exposure rises to 40% when including broader tech and business services, the highest among extended credit markets. We see software volatility as a sector-led reset rather than a start of a macro default cycle.' Similarly, last week in an analyst memo, UBS proposed that investors increasingly want to talk about AI disruption and our tail risk scenario, a rapid severe AI disruption. This is not our...
This 21% exposure thing is important because when John Zito was talking to the journal, or he thought he was talking off the record and was being a little candid, he was saying, you know, I would expect, you know, some of these loans to recover 30 to 40% of the overall loan value. And that just says to me like, okay, if you have 21% exposure, but you can get 30 cents on that, you still recover. You're down 15% overall.
But the goal is not to be down at all.
Yeah. But you're also getting, you know, a lot of these companies have been getting interest, you know, you've been getting some amount of interest back on the loans. So, but again, it doesn't feel like you're getting anything close to a GFC style systemic risk yet, especially when you account for the fund kind of time horizons and the fact they can just say, 'Sorry, you can't get any money out right now.'
So, here's what the Swiss bankers over at UBS had to say. They said, 'What is new is a clearer catalyst. Rapid severe AI disruption. Assuming contagion impacts not modeled in our earlier note, we anticipate US HY, LL, and PC — high yield, leverage loan, and private credit defaults could rise to 3 to 6%, 8 to 10%, and 14 to 15% for private credit respectively.' Blue Owl's updated redemption policy has been compared to BNP Paribas' suspension of redemptions for three funds exposed to US subprime MBS in August of 2007, roughly one year before Lehman Brothers failed for reasons I've stated above. I think that comparison is flawed. The GFC was a systemic crisis driven by leverage, opacity, and reliance on short-term funding. Mortgage credit risk was embedded throughout the financial system and amplified by opaque securitized products like CDO, CDO-squared, and CDO-cubed securities which established tight interdependencies between institutions. Furthermore, the deep interconnectedness of the US financial system rendered the GFC inherently systemic. Home prices were tethered to a mortgage derivatives market that was amplified in value by complex instruments that couldn't easily be valued or understood. The same isn't true of the private credit market, writes Eric Seufert, which comprises relatively straightforward bilateral loans and is largely concentrated within private equity-backed companies and dedicated credit funds. While pockets of stress have emerged, the system is less dependent on short-term funding and less tightly interconnected with core financial infrastructure. While the total value of the private credit market is vastly larger than the high yield energy debt market of the 2014 to 2016 period, my belief is that distress will remain similarly localized and not spill over into the broader economy since there are limited direct risk conduits into the core banking system or short-term funding markets.
So further valuation multiples for PE-backed software companies and especially vertical SaaS companies are compressing for two reasons. One, rising interest rates have increased their debt burden as a result of their floating rate loans. So a lot of these companies were sitting at, you know, with a Fed funds rate of 3%, now it's at six, or I don't know what the Fed funds rate is, but interest rates have basically doubled since the ZERP era and so their debt burden has also increased. And then two, AI and AI-enabled coding tools more specifically jeopardize the defensibility of their businesses. So viewing the composite impact of these two effects as a potential threat to the broader economy ignores the productivity gains posed by AI that imperil vertical SaaS software solutions in the first place. If a company can replace a SaaS vendor with a homemade solution because new AI-enabled engineering tools reduce the time to deployment of a narrow task-specific piece of software to days from months or even years. That's exactly what we've done here. We have a number of point solutions that didn't really exist before but previously we might have gone to some sort of software company and said, oh well, can you do a special version for us or can you change...
Reduces basically like moats.
Yeah, so that same company is not only saving money by no longer paying for that specific point solution but they are also likely accomplishing more across a number of other software development related vectors like clipping in our example. What might be critical here is the hurdle rate in terms of labor to accomplish any task. If one person can build software now with assistance from AI-enabled development tools that previously would have required a team, then building software becomes more accessible generally throughout the economy. For small companies, there is no build versus buy tension. The base cost of SaaS solutions may have simply priced them out of the market completely since hiring a software development team is likewise unattainable. With AI-enabled tools, the cost of building a point solution is reduced dramatically. Certain companies may now benefit from those solutions where before the economics were prohibitive across either building or buying. As Citadel Security notes in a rebuttal it published to a recent macro short thesis that was something big is happening or no, that was the Stratechery prediction, job postings for software developer roles have increased with the availability of AI-enabled development tools. Critically, AI reduces not just the cost of building software but the cost of attempting it, expanding the set of economically viable solutions across the long tail of firms. Not just more software but more bespoke solutions to commercial problems that should result theoretically in better firm-level outcomes that aggregate to greater levels of economic productivity. Let's give it up for better firm-level outcomes. Acemoglu, who won the Nobel Prize in economics last year, speaks to the expansionary effects of technology in his book The Power of Creative Disruption, co-authored with two other authors. 'Automation is thus not in and of itself an enemy of employment. Automation is not an enemy of employment. By modernizing the production process, automation makes firms more competitive, which enables them to win new markets and therefore to hire more employees. This is what we call a productivity effect. This same productivity effect was at work in prior industrial revolutions, those induced by the steam engine and then by electricity, and explains why neither of these revolutions produced the mass unemployment that some had predicted. How can we reconcile this optimistic conclusion with the more pessimistic finding mentioned earlier that automation has a negative effect on employment at the level of commuting zones? One response is to invoke the difficulty of measuring automation and robotization at the commuting zone level as we discussed earlier. A second possible explanation is that firms that do not automate sufficiently end up downsizing their employment, outsourcing their production or simply going under. This would reflect an eviction effect of automation on employment. Firms that invest significantly in new industrial equipment substantially lower their likelihood of going out of business over the following 10 years compared to firms that do not make such an investment. It is thus not automation or manufacturing processes that causes firms to eliminate jobs but rather missing the critical juncture of automation and consequently finding themselves forced to reduce the scope of their activities or even exit the market. In other words, it is through the process of creative disruption that automation can lead to job losses.' In that sense, the vertical SaaS companies acquired by PE firms under low interest rate conditions face headwinds not just from rising rates, but also because it's simply more costly as a result of institutional momentum for them to rebuild their workflows and internal development processes around AI than it is for a company to establish those workflows and development processes in the first place. In other words, established companies face higher costs in adapting to the AI era than do new entrants or companies that are internalizing these tools. De novo and PE-backed firms are fundamentally averse to cost increases.
All right, go subscribe to Mobile Dev Memo. And before we get to our first guest of the show, we got to check in with Jean's, the restaurant in New York. They are weighing in on private credit. They say, 'If you think private credit is a problem, wait until you find out about the annuity businesses they all bought.'
I love that this story has gone so broad that it is now being posted to the Instagram stories of New York restaurants, but that's where we are in this cycle. Anyway, let's move on to Andy Fang, the co-founder of DoorDash, who is waiting for us in the restaurant waiting room. Let's bring him into the shop. Andy, how are you doing?
What's happening? Good to see you again. Welcome back to the show. Oh, wait, actually, we did not have you on the show.
I was first on your show. I'm sorry. Stanley and Tony were both on the show.
That's right. That's right, boys. Is it just the three of you guys, right? Or is there a fourth?
Yeah, it's the three musketeers.
Three musketeers. How's life these days? How's it going?
2026 is good. I mean, lots going on. I mean, I guess, you know, for us, I think AI is like a big focus of mine personally at the company. So, yeah, I mean, recently we just announced...
Yeah. Maybe start with some background on your role, how it's changed over the period of building DoorDash and growing the company. I'm sure it's been a wild ride, but I'd love to hear it from your perspective and then we can go into the acquisition.
Yeah, totally. I mean, let's see, 13 years in the making, I guess. I mean, so I've been always involved on the tech side. So really focused on product and technology at DoorDash. You know, I helped scale up our product and engineering teams for, I don't know, like eight, nine years, first eight, nine years of the company. More after that, we brought on more senior leadership to kind of manage the growing organization and I've been more focused on how do we help DoorDash find the next big bets to help us keep growing. You know, I think whether it's our international businesses, our DashPass subscription program, more recently with the AI stuff. And so I think really just like sinking my teeth into various technology and business problems and figuring out how we can continue to grow the business and find those next seeds.
So talk about how the news today ties into your focus.
Absolutely. So yeah, I was just briefly mentioning, so AI is like something that we're really trying to figure out at DoorDash, how we can incorporate it into our business. And so earlier today we just announced that we acquired a company called Metis. They are an applied AI research lab. They went through YC and we met with them, I want to say like six or seven months ago. And I think we were thinking about how to...
Was that intentional? Be honest.
I don't know, I was like...
It really just pops into your brain. It's such a weird...
I just, oh my god, I'm such an old head. I was reading through some Wall Street Journal article and I kept saying the two numbers that shall not be named together. And Jordi was like, 'Are you doing this intentionally?'
Absolutely. So yeah. So then September is when we first started talking to them. And we were trying to figure out how to productionize AI use cases at DoorDash. And so really, I mean, we kind of spread in terms of talking to a bunch of different startups. There's a lot of startups like trying to figure out how to deploy AI for like the enterprise use case and we just decided to be really open-minded about it and partner with a bunch of them. But Metis really stuck out to us. I mean, we were working with them since September. There are some really cool use cases around agentic commerce and physical intelligence that I can talk about a bit that we're working with them on that we're super excited to accelerate with them joining the company. So, yeah. I mean, they're like a bunch of cracked 22-year-old like AI kids and they're like super high hustle, very knowledgeable about what's going on in the space and, you know, we just want to inject a lot of that energy into how we're trying to innovate here.
Yeah. Maybe take us through some of the history of AI at DoorDash because I imagine you were doing recommendation systems very early, matching, all sorts of stuff. And then when the LLM revolution happens, like the obvious bolt-on like chat box went into a lot of apps. Did that happen? And then everyone starts using coding agents and IDEs. So what have been the key turning points just internally at DoorDash in AI adoption?
Yeah, I mean, so I think to your point, I think there's two themes in terms of how we're trying to think about it. I think one is like employee enablement or like productivity. I think obviously you guys are very well aware of like since November basically, I think there's been an explosion especially on the coding side in terms of people being much more productive. I'd say I think at DoorDash our evolution in terms of AI on like the product facing side has been a longer journey in the sense that I think we've been trying to figure out how this all fits into our marketplace and how this would benefit our customers. Like to your point, I think a lot of people, ourselves included, were just trying to bolt on some sort of chat box just to see how it would play in. And I think one, when we did that a couple years ago, the LLMs were not nearly as smart. But two, I think we just need to be more thoughtful, like how do we integrate this into the experience? And I think agentic commerce is something that a lot of people have been talking about and trying to figure out. And for us, I would say like agentic commerce is basically figuring out how we can leverage this technology to make it actually easier for restaurants to run their business and for customers to find what they want on DoorDash. You know, I think one of the big problems that we've had on DoorDash and what people tell me all the time when I talk to them is like, 'Look, Andy, like you guys got a ton of selection. I have no idea what to order.' You know, it's like there's so much, you know, so many different restaurants and all these different cuisines. Like I'm having like decision paralysis right now. And so I think like that's a very interesting problem where we feel like AI can be beneficial in helping you find what you want and give you confidence in that as a customer.
One of the greatest luxuries or first world problems in life is decision paralysis on what to eat. Comes for everyone.
Yeah. Have you been seeing folks like vibe code sort of like plugins or add-ons or features that then you can potentially use to adopt or roll into the actual roadmap? I swear I saw someone post something about this where they vibe coded basically a Chrome plugin that would filter DoorDash for vegetarian options and it was like a pretty mature functionality and they just went through like a series of prompts and got to like something that they were very satisfied with. But that feels like they're sort of doing the R&D on their end and then once it's working you can actually bring that back. I'm wondering about this like primordial explosion of people building custom software and then the pipeline to actually bring that into the business as new features.
That's interesting because I think there's some... I haven't heard of that one specifically to be honest, but I think there has been a lot of interest like, hey, can you figure out a way to enable whatever my agent or whatever software I'm building to integrate into the DoorDash ecosystem? Like I've gotten pings from founders be like, yo, when's the DoorDash CLI coming out? And I'm like, hold on, like we need to figure out what that means. And I think I would say, you know, and we also did something with OpenAI in terms of like being an early partner with our apps ecosystem and stuff. So we're trying to figure out how we want to play in this space to be honest.
So just walk me actually through that OpenAI partnership because like wouldn't that be the CLI? Like if I have Codex that's a CLI and then if you have a partnership with OpenAI, can't Codex CLI talk to DoorDash then or is there some break in the chain that needs to be resolved?
I think it's more like in ChatGPT, it's not really accessible through Codex. So I have to tell Codex to go use ChatGPT, open a web browser, click on the ChatGPT box. It is weird like these un-hobblings are so intuitive and then you realize that okay there's actually some sort of business reason or whatnot. Yeah.
Yeah. I mean, I think a lot of people are just trying to throw stuff at the wall and see what sticks. So I think one thing that we've been finding really like getting a lot of traction especially with like what we're working with Meta on is like I think there's some actual agentic workflows, like customer-facing workflows that we're working on that like just like us playing around with it the past couple weeks like we legitimately think this is going to be something that's going to be beneficial to customers and help you find what you want and help you order like...
How do you think about even positioning these products? Because like obviously like broad public opinion on AI is not great. Is this the kind of thing I'm assuming you're just like trying to deliver a great experience and not, you know, slap like DoorDash AI, the thing you've been waiting for.
Yeah, exactly. I think we want to introduce it in a way that's actually beneficial to like conversion or like you finding what you want and being satisfied with it. And so that's why we've taken our time in some ways in terms of like rolling something out that we want to launch more broadly. But I think some of the challenges we've had historically is like these agents are not very smart at like finding you what you want. This was like a while ago, right? Obviously things have gotten a lot smarter. The question is how do we leverage the intelligence that you know a lot of like harness building context engineering type stuff. But it's also like how do you build a great experience around it? Like you know there's like obviously LLMs, they're super smart. Sometimes they like are slower with their reasoning, but we're just trying to find that balance. And like I think to your point, it's like how do we create an interesting experience that doesn't just feel like we're slapping something AI on?
Are you going to do anything with voice at all? Like I can imagine somebody in the DoorDash app saying like, 'Hey, I want Chinese food. I want to avoid...'
Tells you that they integrated with Siri a decade ago.
Yeah. Yeah. We do have some voice modality stuff, but I think I would say that's a little bit more like just a modality in terms of how you place your order and like I guess if it's more natural language with the technology like that's a...
Yeah, I'm saying like yeah, it's a natural language. You're kind of searching over all the offerings in your area and then it's like potentially putting together like a cart for you or an order you can check out. A couple more things. How does Metis tie into the physical AI autonomy? You mentioned something about that earlier.
Yeah. So, I mean, I think DoorDash wants to play at the frontier of physical intelligence. I think that's something we talked about for a while. Obviously, our...
You guys have your cute little robots.
Exactly. The Dot delivery robots. So, my co-founder Stanley, who has been working on that full-time, like we're super excited to deploy those more broadly. But, you know, we have a large courier network. We have a large support network where we're collecting billions of data points in the physical world every day and it's like how can we leverage that to actually push forward what's possible with robotics and physical intelligence. I think, you know, there's some interesting conversations we've had with various startups and labs on how we collaborate to push forward the frontier together and I think Metis has some great experience doing that type of stuff in the expert data space. I mean, you know, Arian, the CEO, is like a former Meta, so I think there's some really interesting stuff there that we're interested to explore and yeah, we're excited to accelerate some of that together with them.
Awesome. Matt from Tempo is coming on later today. Tempo is launching today. I know you guys were an early design partner. Talk about kind of that decision to get involved.
It's funny you bring that up because Matt literally texted me before this, hey, we're going to be on the show together. I was like, sick. So yeah, we're excited. I mean, I'll let Matt talk more about that, but I think some of actually what Matt's working on is going to be synergistic with the agentic commerce stuff we're working on potentially, but essentially like we are very excited to just partner with companies like Tempo to see how stablecoins can be disruptive and innovative in our space. More to come in terms of what exactly our pilot use case is going to be. But yeah, we've really enjoyed working with them and, you know, Matt's been great. We've known Matt for a while. So...
I want to ask one more about agentic commerce. I'm trying to think about like the product that I might actually want. I understand like when most people say agentic commerce, it's like open up ChatGPT and fire off a query, say order me a burrito, and it goes through DoorDash and you handle the rails. But I'm thinking more about like the next version of that. And maybe it's something like, okay, I want lunch every day, but I need you to check my calendar because the show ends sometimes at 2, sometimes at 2:30. That information's on my calendar, but also sometimes I'm leaving to go to the airport and I'm flying, so I'm not available. And I could potentially like encode all these preferences. I have favorite things at different restaurants. I'd like some variety, but some consistency. Maybe you should fire off a text message to me and let me know that I can opt out of today's order. And I feel like that's like the system that someone would vibe code if you gave them the CLI, but that should probably just live within DoorDash. So, how do you think about that sort of operating at a higher level, like encoding all of your preferences and then putting your DoorDash experience like basically on autopilot?
Totally. I mean, I think one thing we say to ourselves is like if someone's building DoorDash today, it would not look like DoorDash, you know, like it would be... I think and when we think about agents versus like just like chatbots, like agents in an ideal world, they're not just telling you stuff, they're actually doing stuff for you. And I think I would say from a consumer-facing use case perspective, in my opinion, I think besides the chatbots, I think OpenClaw is like kind of the closest thing we've seen to something that is like where agentic commerce is heading towards. And so we're very inspired by that type of... I don't know if use case is the right...
word, but that model and that paradigm of like agents doing stuff on behalf of you. And I think your example kind of hit it on the head, is like, I don't want to just tell me stuff, do stuff for me. And I think we have a ton of information on your preferences. We have a ton of real-time information on what's going on logistics-wise in your neighborhood and your city, like real-time inventory information in restaurants and retailers. And so we can create some magical experiences there. And again, for us, it's not just about slapping this technology on for the sake of it, but it's actually stuff that restaurants would really appreciate because we can help them run their business better.
Like a concrete example on the restaurant side, and I'll get back to your use case in a second, but on the restaurant side, it's like, I don't know how to price my pad thai in my neighborhood, you know, or what deals or promotions should I be running right now, you know? Because I just want to figure out how to grow my sales, or help me figure out how to, you know, the lulls between lunch and dinner, help me figure out how to get more demand there. And then on the customer side, I think to your example, it's like, you know, I don't want to think about ordering my lunch to the office every day or my weekly groceries.
Which is a crazy thing to say because you basically invented the 'I'm hungry' button on your phone, and it's literally like two clicks for me when I use DoorDash, and I'm like, we still... I don't want the horse anymore. Give me the car.
I didn't have lunch yesterday. We actually messed it up somehow. We at the office forgot to order. Yeah, we forgot to order actually. No, I did have a salad, but it was like hours later and I figured it out.
How rare is it to be 13 years into a business that has been as successful as DoorDash to still have three founders in the business?
Yeah. It feels like I was trying to come up with an example. Narrative violation at this stage.
That's interesting. I did not expect that. But I don't take that for granted at all. I mean, in fact...
They secretly hate each other, right? But there's some weird contract where if any of you leave, the whole company shuts down, so you're locked to each other, handcuffed on the Titanic together. No, I'm messing with you.
I think I'm very grateful for the dynamic and the relationship we have. I mean, I think Tony is a one-of-one CEO, and I think we have a really good dynamic together, and we put a lot of trust in each other to just do the right thing.
Have you guys always had the same roles basically, and sort of known, okay, I'm not going to step on this person's toes in this particular region because then there's mutual trust, as opposed to a lot of founding teams, they get in trouble when there's three people that want to be CEO and they're fighting over it constantly. I mean, that's usually what happens.
Oh yeah, it's true. I mean, I would say we delineated our responsibilities pretty early on, and I think Tony's role has stayed consistent. I mean, CEO is CEO, but I think Stanley and I, our roles have evolved around both how to stay on the forefront of innovation, but also just making sure we're thinking about what are the most important problems to solve. But no, I think we have a really good relationship. I think it's high trust. We're getting dinner tonight. We try to catch up every so often. But yeah.
Door-dashing dinner tonight?
No, we're gonna go out. I don't know where we're going out, but we picked a spot.
Yeah, it's a special relationship. I did not expect that, but yeah, it's right there.
No, it's cool. Last question. After the very silly Catrini piece, have you guys thought about doing a hackathon and getting your best engineers to be like, 'Hey, just vibe code DoorDash. Make the agent-native DoorDash.' Because I think it would actually be really interesting to just actually try to do it and see.
You know, it's funny you bring that up because that Satrina report blew up out of nowhere. And I don't know. I mean, honestly, we didn't react too much to it. I mean, I tweeted something in response to it just in terms of my own thoughts there, but I would say in general...
Two words: network effects.
Yeah. I mean, I would say in general, we just try to focus on growing our business and focusing on customers, but I don't know. Like we try to get people to vibe code a bunch of different use cases, but not try to vibe code DoorDash.
Yeah, there's more to it. More to it. A Bits and Adams company. Well, awesome. Great. Great to meet you, Andy. Next time, bring all three.
Yeah. We'd love to have all three of us.
